AWS Bedrock Pricing Explained: What You Actually Pay Per Token

AI This article was created with the help of AI.

How does Bedrock charge for tokens?

Amazon Bedrock meters foundation model usage primarily on a pay-as-you-go basis, charging separately for input and output tokens. Because prompt processing and token generation demand different computational resources on underlying GPU clusters, AWS assigns distinct prices to each stream (Last verified: July 2026). Evaluating AI compute costs using a single blended token average misleads infrastructure capacity planning, as input-heavy Retrieval-Augmented Generation (RAG) workloads exhibit vastly different financial profiles than output-heavy code generation or agent reasoning chains.

Understanding Input vs. Output Token Economics

Input tokens represent the text, prompt templates, and context windows ingested by the model, whereas output tokens represent the generated response. On AWS Bedrock, output tokens carry a substantial price premium over input tokens across virtually every provider catalog. For example, Anthropic Claude 3.5 Sonnet on-demand pricing is set at $6.00 per 1M input tokens and $30.00 per 1M output tokens. Because generation is priced well above ingestion, an application producing lengthy technical documentation or detailed JSON outputs will accumulate costs significantly faster than one processing long context windows with concise classification outputs.

To mitigate high on-demand rates for non-real-time workloads, AWS offers a batch inference tier for select foundation models. Batch mode processes asynchronous request queues at 50% less than on-demand rates, which puts Claude 3.5 Sonnet at $3.00 per 1M input tokens and $15.00 per 1M output tokens. However, batch processing introduces latency delays, making it unsuitable for interactive user applications or real-time API integrations.

When does provisioned throughput beat on-demand?

For high-throughput enterprise applications requiring guaranteed inference capacity and sub-second latency, relying strictly on on-demand token metering creates both cost and rate-limit risks. AWS addresses this through Provisioned Throughput, where teams reserve dedicated model capacity allocated in Model Units (MUs). Instead of paying per token, provisioned throughput charges a fixed hourly fee per model unit under term commitments.

Calculating the Break-Even Crossover

Provisioned Throughput pricing varies by model family, architecture size, and commitment length. For Cohere Command on Bedrock, an uncommitted provisioned model unit costs $49.50 per hour. Committing to a 1-month term lowers the rate to $39.60 per hour per model unit, while a 6-month term reduces the commitment to $23.77 per hour. Deciding whether to transition from on-demand pay-per-token metering to provisioned throughput requires precise workload profiling.

1. Establish baseline continuous token volume: Calculate total daily input and output token consumption during steady-state operations.
2. Determine peak concurrent request requirements: Model units enforce hard limits on concurrent transactions per second (TPS). If peak traffic exceeds on-demand rate limits, provisioned capacity becomes mandatory regardless of baseline cost.
3. Calculate monthly fixed commitment: Multiply the committed hourly model unit rate by the hours in the billing month to get the fixed monthly baseline cost per MU.
4. Map against on-demand crossover point: Compare that fixed monthly commitment against what the same token volume would cost under standard per-token rates. If monthly on-demand spend consistently exceeds the fixed MU fee, provisioned throughput delivers lower unit costs per token.

The primary engineering risk with Provisioned Throughput lies in utilization waste. Provisioned model units charge continuously 24 hours a day, 7 days a week. If application traffic drops significantly overnight or during weekends, the effective cost per generated token spikes sharply, offsetting the nominal discount achieved during peak hours.

Which Bedrock models carry a price premium?

Model pricing on AWS Bedrock spans multiple orders of magnitude depending on parameter count, provider licensing, and deployment architecture. Flagship reasoning models and frontier architectures command the highest baseline token rates, while smaller task-specific models offer drastically lower unit economics.

Geographic Routing and In-Region Premiums

In addition to base model capability tiers, geographic deployment options introduce substantial price variations. AWS provides Global cross-region routing to dynamically distribute request traffic across global infrastructure pools, often maintaining lower baseline pricing. However, European enterprises operating under strict data protection frameworks frequently require in-region inference pinned strictly to local European data centers.

As the table shows, selecting local European hosting in Stockholm over US East regions raises DeepSeek v3.2 token rates by roughly a fifth on both the input and the output side. That regional uplift reflects the higher operational, power, and compliance costs of European data center infrastructure. Engineering teams conducting cost per million tokens evaluations must account for these regional differences when modeling production deployment budgets.

Do Bedrock costs include data transfer?

A common oversight when budgeting for foundation model APIs is assuming that per-token prices cover end-to-end network delivery. On AWS Bedrock, token rates cover GPU execution and model processing only. Moving data into, across, or out of the AWS environment incurs separate network line items.

Egress Fees and Multi-Region Architectures

While data ingress into AWS Bedrock is free, external network egress and cross-region routing are subject to standard AWS data transfer charges:

• Internet Data Egress: Transmitting model outputs, generated images, or high-volume streaming responses out of AWS to external client applications costs $0.09 per GB after initial free tier allowances.
• Cross-Region Data Transfer: When using cross-region inference or calling Bedrock endpoints located in a different AWS region than the primary application server (e.g., calling an inference endpoint in US East from an EC2 instance in Frankfurt), AWS applies its standard per-GB inter-region transfer charge on top of token costs.
• AWS PrivateLink Charges: Deploying private VPC endpoints for Bedrock to prevent traffic from crossing the public internet adds hourly endpoint fees plus per-GB data processing surcharges.

For text-only interactions, data transfer fees represent a modest fraction of overall spend. However, for multimodal applications processing high-resolution images, audio streams, or massive document batches, cloud egress fees accumulate rapidly, and providers differ in whether they meter network transfer at all.

What drives a surprise Bedrock bill?

When organizations experience unexpected spikes in their AWS Bedrock invoices, the core foundation model token line is rarely the sole culprit. Instead, secondary platform features and orchestration tools built on top of the model endpoints add stacked per-use charges that multiply the baseline compute spend.

Stacking Charges Beyond Token Lines

Building end-to-end generative AI applications on Bedrock typically involves orchestrating retrieval, safety filtering, and agent loops. Each of these capabilities operates under an independent metering model:

• Bedrock Guardrails: Content filtering, PII masking, and topic policy enforcement are billed per evaluation, and one text unit contains up to 1,000 characters, so a longer input is charged as several units. Crucially, Guardrails evaluate both the incoming user prompt and the outgoing model completion, effectively doubling the evaluation charge per request.
• Bedrock Knowledge Bases: Managed RAG workflows combine multiple billing layers. Beyond the underlying vector storage costs (such as OpenSearch Serverless compute units), retrieving contextual passages incurs embedding model token fees plus separately metered Bedrock Data Automation charges for parsing ingested documents.
• Bedrock Rerank: Re-ranking search results before passing context to the LLM is billed per query, where a single query can contain up to 100 document chunks; larger result sets are metered as multiple queries.
• Bedrock Agents: Autonomous agents execute multi-step reasoning loops. Each step invokes the underlying LLM to decide tool calls, so a single user query consumes several times the input and output tokens of a direct single-turn prompt.

When these services operate concurrently, a single agent interaction that performs safety filtering, context retrieval, document re-ranking, and multi-turn reasoning can cost five to ten times more than the raw LLM token rate alone.

How does model availability impact EU data residency?

For European engineering teams and enterprise architects, pricing cannot be evaluated separately from regulatory compliance. Aligning model selection with European data sovereignty requirements introduces technical and geographic constraints on AWS.

Cross-Border Data Transfer and the CLOUD Act

Model availability across AWS regions is non-uniform. Newer model releases, specialized model variants, or specific provider integrations frequently launch first in US East (N. Virginia or Oregon) before deploying to European regions like eu-central-1 (Frankfurt) or eu-west-1 (Ireland).

• Model Parity Gaps: European deployment regions often lack immediate access to the latest frontier model checkpoints, forcing teams to either delay features or route payload traffic to US data centers.
• Cross-Region Failover Exposure: Enabling cross-region routing or automated failover to maintain high availability can dynamically redirect inference requests across international borders during regional outage events or capacity bottlenecks.
• Jurisdictional Frameworks: Because AWS is a US-headquartered corporation, it remains subject to the US CLOUD Act. Under European data protection standards and Schrems II jurisprudence, US legal extra-territorial reach introduces persistent legal risk, even when workloads are hosted within European geographic borders.

Navigating these compliance challenges requires understanding data transfer risks and ensuring that model endpoints guarantee strict, legally binding data residency without hidden cross-border routing mechanisms.

How does Bedrock compare to open-model APIs?

Conflict of interest disclosure: Lyceum publishes this guide and operates an EU-sovereign AI infrastructure platform competing in serverless inference.

Comparing Frontier Closed Models vs. EU-Native Open Weights

Evaluating AWS Bedrock against open-weight serverless alternatives highlights a growing divergence in AI unit economics. Proprietary closed-source models hosted on Bedrock carry substantial margins. Paying $6.00 per 1M input tokens and $30.00 per 1M output tokens for Claude 3.5 Sonnet represents a major operational commitment for enterprise production workloads.

By contrast, frontier open-weight models deployed on specialized European inference platforms deliver comparable reasoning, coding, and multilingual capabilities at a fraction of the cost. On EU-hosted serverless endpoints, engineering teams access models like DeepSeek V4 Pro at $1.75 per 1M input and $3.50 per 1M output, or high-speed models like DeepSeek V4 Flash at $0.15 per 1M input and $0.30 per 1M output.

Whether teams rely on serverless endpoints or scale up to dedicated GPU inference, the platform behind this guide runs 30+ open-weight models behind a drop-in OpenAI-compatible API hosted in eu-north1, with per-token serverless billing and European data residency. The live pricing page and its savings calculator let you compare that rate card against your current Bedrock spend.